Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

test(cuda): group tier regression rules with related tests

ZacharyZcR39adcc69f909
diff --git a/c/Makefile b/c/Makefile
index a36104de..e01e9760 100644
--- a/c/Makefile
+++ b/c/Makefile
@@ -1834,6 +1834,10 @@ tests/test_qwen36_tier_int8$(EXE): tests/test_qwen36_tier_int8.c tests/qwen36_fa
tests/test_qwen36_tier_multidev$(EXE): tests/test_qwen36_tier_multidev.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h
$(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)
+# Budget separate gate/up/down scale allocations at their own size classes.
+tests/test_qwen36_tier_scale_budget$(EXE): tests/test_qwen36_tier_scale_budget.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h
+ $(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)
+
# Same fake backend, single device: proves qt_shutdown returns (under an
# alarm(10) watchdog) instead of hanging forever when a group is still open
# and an LFRU swap is parked waiting for cv_take (#1340).
@@ -2180,7 +2184,3 @@ tests/test_kv_prefix$(EXE): tests/test_kv_prefix.c kv_prefix.h
tests/test_kimi_request_state$(EXE): tests/test_kimi_request_state.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h
$(CC) $(NOCUDA_CFLAGS) tests/test_kimi_request_state.c $(VK_OBJ) -o tests/test_kimi_request_state$(EXE) $(NOCUDA_LDFLAGS)
-
-# Budget separate gate/up/down scale allocations at their own size classes.
-tests/test_qwen36_tier_scale_budget$(EXE): tests/test_qwen36_tier_scale_budget.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h
- $(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)